Overview & result
0xShikhar presents JEV Fuse as a governed execution layer between Jev and agent tooling such as Claude Code, MCP, and SDKs. It converts raw probabilistic decisions into ALLOW, ASK, or DENY and adds deterministic shell guards, abstention around ambiguous scores, deduplication, and an audit trail.
From the original post
Been messing around with @typesafeai JEV + agent tooling from last weekend and ended up building a little piece of infrastructure around it.
The weird thing about fast decision models is that you can get a probability in milliseconds.. but a probability isn't an execution policy.
So I built JEV Fuse ⚡
A governed execution layer between decision models and agent tooling (Claude Code, MCP, SDKs) that turns raw decisions into deterministic actions:
ALLOW / ASK / DENY
Under the hood:
• Deterministic shell AST guards
• Deadband abstention around ambiguous decisions
• In-memory caching + singleflight deduplication
• Persistent SQLite audit trail
the idea: don't let a probabilistic output directly touch a production tool.
its a drop-in proxy and completely open source👇
How Jev fits in the loop
- The decision model receives bounded state and returns a typed probability.
- JEV Fuse applies deterministic AST guards and a deadband policy before execution.
- The agent runtime receives ALLOW, ASK, or DENY and the decision is recorded for audit.
How to reproduce
- Clone the verified repository and follow its documented installation path.
- Send a typed decision through the proxy and inspect the resulting policy action.
- Review the audit record and deadband behavior before connecting an irreversible tool.
Why this build matters
The project draws the boundary between deciding and executing: a probability is not allowed to touch a production tool without a policy layer.
Limitations
- The post describes an open-source infrastructure project but does not provide an independent benchmark.
- Production policy thresholds and the set of guarded tools still require project-specific review.